Russian AI Expert to Flag Errors in Construction Project Documents
Homegrown AI model is designed to check compliance with regulatory requirements
Glavgosexpertiza has begun beta testing its own language model. The system analyzes construction project documentation and automatically identifies non-compliance with regulatory standards.
Digital Review
Vadim Andropov, First Deputy Head of Glavgosexpertiza of Russia, presented the new development at a meeting of the Federation Council’s section on digital transformation in construction. The discussion focused on shifting construction services to electronic formats and the practical impact of that transition.
According to Dmitry Kuzmin, the move to electronic document management and remote interaction has already reduced project review timelines and increased transparency. For regional authorities, this directly affects the pace of construction.
Unified Platform
Work on the Unified Digital Platform for Expert Review has accelerated document processing severalfold. The time required for an initial review has fallen from 10 days in 2016 to two days in 2025.
Offline AI Model
Glavgosexpertiza has deployed its own local language model. It operates in isolation from external networks, helping ensure data security and control. Specialists further trained the system using RAG (Retrieval-Augmented Generation) technology to improve the accuracy of responses based on relevant regulatory documents. In practice, the model does not rely solely on prior training but also searches a document database, retrieves relevant fragments and incorporates that context when generating responses.
According to Andropov, the system can achieve up to 95 percent accuracy through tuning and specialization.
Who Is Responsible for AI?
The rollout of AI raises questions about accountability, including so-called model hallucinations and potential data leaks. Officials say the sector needs a dedicated regulatory framework.
In the longer term, Glavgosexpertiza plans to expand its data-driven approach. Experts will increasingly work as data engineers, and the review process itself could evolve into a service model for developers, potentially offering “one-click review” functionality.
